[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/libcudacxx/include/cuda/__execution/guarantee.h

@@ -0,0 +1,94 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CUDA___EXECUTION_GUARANTEE_H
+#define __CUDA___EXECUTION_GUARANTEE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+// Base class of all guarantees that can be passed to cuda::execution::guarantee.
+//
+// A guarantee is a promise that the caller makes to an algorithm about its input or the problem being solved (e.g. an
+// upper bound on the total number of items). Algorithms may exploit guarantees to select faster code paths or smaller
+// intermediate types, but they are never required to. This is the dual of cuda::execution::__requirement, which
+// describes a property that the caller demands from the algorithm. Unlike requirements, guarantees may be stateful,
+// i.e. they may carry a runtime value.
+class __guarantee
+{};
+
+// Query used by algorithms to extract the environment of bundled guarantees from an environment assembled by
+// cuda::execution::guarantee.
+struct __get_guarantees_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_guarantees_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)), "Guarantees must be queryable without throwing");
+    return __env.query(*this);
+  }
+
+  // Answering the forwarding_query query with true makes __get_guarantees_t a forwarding query: environment adaptors
+  // pass it through to the environments they wrap, so guarantees stay visible to the algorithm no matter how deeply
+  // the environment carrying them is nested.
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_guarantees = __get_guarantees_t{};
+
+//! @brief Bundles a pack of guarantees into an environment that can be passed to device-wide parallel algorithms.
+//!
+//! @param __guarantees The guarantees to bundle. Each guarantee must be nothrow copy constructible. It is stored by
+//! value, preserving any runtime state it carries.
+//! @return An environment carrying the bundled guarantees.
+// The returned property is keyed by __get_guarantees_t so that individual guarantees are only visible to algorithms
+// through the guarantees environment, mirroring how cuda::execution::require exposes requirements. The key is spelled
+// __get_guarantees_t{} rather than reusing the __get_guarantees global: odr-using the global inside this inline
+// host-device function would reference a per-TU internal-linkage object in device passes (see _CCCL_GLOBAL_CONSTANT).
+template <class... _Guarantees>
+[[nodiscard]] _CCCL_NODEBUG_API constexpr auto guarantee(_Guarantees... __guarantees) noexcept
+{
+  static_assert((::cuda::std::is_base_of_v<__guarantee, _Guarantees> && ...),
+                "Only guarantees can be passed to guarantee");
+  static_assert((::cuda::std::is_nothrow_copy_constructible_v<_Guarantees> && ...),
+                "Guarantees must be nothrow copy constructible");
+
+  ::cuda::std::execution::env<_Guarantees...> __env{__guarantees...};
+
+  return ::cuda::std::execution::prop{__get_guarantees_t{}, __env};
+}
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_GUARANTEE_H